feat: Cache-aligned bytecode storage

crates/plotnik-cli/src/commands/dump.rs

@@ -82,7 +82,7 @@ pub fn run(args: DumpArgs) {
         query.emit().expect("bytecode emission failed")
     };
 
-    let module = Module::from_bytes(bytecode).expect("module loading failed");
+    let module = Module::load(&bytecode).expect("module loading failed");
     let colors = Colors::new(args.color);
     print!("{}", dump(&module, colors));
 }

crates/plotnik-cli/src/commands/infer.rs

@@ -96,7 +96,7 @@ pub fn run(args: InferArgs) {
         }
         query.emit().expect("bytecode emission failed")
     };
-    let module = Module::from_bytes(bytecode).expect("module loading failed");
+    let module = Module::load(&bytecode).expect("module loading failed");
 
     // Emit TypeScript types
     let void_type = match args.void_type.as_deref() {

crates/plotnik-cli/src/commands/run_common.rs

@@ -194,7 +194,7 @@ pub fn prepare_query(input: QueryInput) -> PreparedQuery {
     }
 
     let bytecode = emit_linked(&query).expect("emit failed");
-    let module = Module::from_bytes(bytecode).expect("module load failed");
+    let module = Module::load(&bytecode).expect("module load failed");
 
     let entrypoint = resolve_entrypoint(&module, input.entry);
     let tree = lang.parse(&source_code);

crates/plotnik-lib/src/bytecode/aligned_vec.rs

@@ -0,0 +1,108 @@
+//! 64-byte aligned storage for bytecode.
+//!
+//! Bytecode sections are 64-byte aligned internally. For this alignment to be
+//! meaningful at runtime, the buffer itself must start at a 64-byte boundary.
+//! Standard `Vec<u8>` provides no alignment guarantees for `u8`.
+
+use std::ops::Deref;
+
+/// Alignment for bytecode buffers (matches `SECTION_ALIGN`).
+pub const ALIGN: usize = 64;
+
+/// 64-byte aligned block for bytecode storage.
+#[repr(C, align(64))]
+#[derive(Clone, Copy)]
+struct Block([u8; 64]);
+
+/// Immutable 64-byte aligned byte storage.
+///
+/// Uses `Vec<Block>` internally — Vec guarantees element alignment,
+/// so the data starts at a 64-byte boundary. No custom allocator needed.
+pub struct AlignedVec {
+    blocks: Vec<Block>,
+    len: usize,
+}
+
+impl AlignedVec {
+    /// Copy bytes into aligned storage.
+    pub fn copy_from_slice(bytes: &[u8]) -> Self {
+        if bytes.is_empty() {
+            return Self {
+                blocks: Vec::new(),
+                len: 0,
+            };
+        }
+
+        let num_blocks = bytes.len().div_ceil(64);
+        let mut blocks = vec![Block([0; 64]); num_blocks];
+
+        // Copy block by block to stay safe
+        for (i, chunk) in bytes.chunks(64).enumerate() {
+            blocks[i].0[..chunk.len()].copy_from_slice(chunk);
+        }
+
+        Self {
+            blocks,
+            len: bytes.len(),
+        }
+    }
+
+    /// Read a file into aligned storage.
+    pub fn from_file(path: impl AsRef<std::path::Path>) -> std::io::Result<Self> {
+        let bytes = std::fs::read(path)?;
+        Ok(Self::copy_from_slice(&bytes))
+    }
+
+    /// Number of bytes stored.
+    pub fn len(&self) -> usize {
+        self.len
+    }
+
+    /// Check if empty.
+    pub fn is_empty(&self) -> bool {
+        self.len == 0
+    }
+
+    /// View as byte slice.
+    pub fn as_slice(&self) -> &[u8] {
+        if self.blocks.is_empty() {
+            return &[];
+        }
+        if self.len > self.blocks.len() * 64 {
+            panic!(
+                "AlignedVec invariant violated: len {} exceeds capacity {}",
+                self.len,
+                self.blocks.len() * 64
+            );
+        }
+        // SAFETY: Block is repr(C) with only [u8; 64], so pointer cast is valid.
+        // We only expose `len` bytes, which were initialized in copy_from_slice.
+        unsafe { std::slice::from_raw_parts(self.blocks.as_ptr() as *const u8, self.len) }
+    }
+}
+
+impl Deref for AlignedVec {
+    type Target = [u8];
+
+    fn deref(&self) -> &[u8] {
+        self.as_slice()
+    }
+}
+
+impl Clone for AlignedVec {
+    fn clone(&self) -> Self {
+        Self {
+            blocks: self.blocks.clone(),
+            len: self.len,
+        }
+    }
+}
+
+impl std::fmt::Debug for AlignedVec {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("AlignedVec")
+            .field("len", &self.len)
+            .field("aligned", &(self.blocks.as_ptr() as usize).is_multiple_of(ALIGN))
+            .finish()
+    }
+}

crates/plotnik-lib/src/bytecode/aligned_vec_tests.rs

@@ -0,0 +1,76 @@
+use super::aligned_vec::{AlignedVec, ALIGN};
+
+fn is_aligned(ptr: *const u8) -> bool {
+    (ptr as usize).is_multiple_of(ALIGN)
+}
+
+#[test]
+fn alignment_guarantee() {
+    let data: Vec<u8> = (0..100).collect();
+    let vec = AlignedVec::copy_from_slice(&data);
+    assert!(is_aligned(vec.as_ptr()));
+}
+
+#[test]
+fn copy_from_slice() {
+    let data = [1u8, 2, 3, 4, 5];
+    let vec = AlignedVec::copy_from_slice(&data);
+
+    assert!(is_aligned(vec.as_ptr()));
+    assert_eq!(&*vec, &data);
+}
+
+#[test]
+fn empty_slice() {
+    let vec = AlignedVec::copy_from_slice(&[]);
+    assert!(vec.is_empty());
+    assert_eq!(vec.len(), 0);
+    assert_eq!(vec.as_slice(), &[] as &[u8]);
+}
+
+#[test]
+fn clone_preserves_alignment() {
+    let data: Vec<u8> = (0..100).collect();
+    let vec = AlignedVec::copy_from_slice(&data);
+    let cloned = vec.clone();
+
+    assert!(is_aligned(cloned.as_ptr()));
+    assert_eq!(&*cloned, &*vec);
+}
+
+#[test]
+fn deref_to_slice() {
+    let vec = AlignedVec::copy_from_slice(&[10, 20, 30]);
+
+    let slice: &[u8] = &vec;
+    assert_eq!(slice, &[10, 20, 30]);
+    assert_eq!(vec[0], 10);
+    assert_eq!(vec[2], 30);
+}
+
+#[test]
+fn large_data() {
+    let data: Vec<u8> = (0..10_000).map(|i| (i % 256) as u8).collect();
+    let vec = AlignedVec::copy_from_slice(&data);
+
+    assert!(is_aligned(vec.as_ptr()));
+    assert_eq!(&*vec, &data[..]);
+}
+
+#[test]
+fn partial_block() {
+    let data: Vec<u8> = (0..37).collect();
+    let vec = AlignedVec::copy_from_slice(&data);
+
+    assert_eq!(vec.len(), 37);
+    assert_eq!(&*vec, &data[..]);
+}
+
+#[test]
+fn exact_block_boundary() {
+    let data: Vec<u8> = (0..128).map(|i| i as u8).collect();
+    let vec = AlignedVec::copy_from_slice(&data);
+
+    assert_eq!(vec.len(), 128);
+    assert_eq!(&*vec, &data[..]);
+}

crates/plotnik-lib/src/bytecode/mod.rs

@@ -2,6 +2,7 @@
 //!
 //! Implements the binary format specified in `docs/binary-format/`.
 
+mod aligned_vec;
 mod constants;
 mod dump;
 mod effects;
@@ -16,6 +17,8 @@ mod nav;
 mod sections;
 mod type_meta;
 
+pub use aligned_vec::AlignedVec;
+
 pub use constants::{
     MAGIC, MAX_MATCH_PAYLOAD_SLOTS, MAX_PRE_EFFECTS, SECTION_ALIGN, STEP_SIZE, VERSION,
 };
@@ -56,6 +59,8 @@ pub use ir::{
     TrampolineIR,
 };
 
+#[cfg(test)]
+mod aligned_vec_tests;
 #[cfg(test)]
 mod effects_tests;
 #[cfg(test)]

crates/plotnik-lib/src/bytecode/module.rs

@@ -7,6 +7,7 @@ use std::io;
 use std::ops::Deref;
 use std::path::Path;
 
+use super::aligned_vec::AlignedVec;
 use super::header::{Header, SectionOffsets};
 use super::ids::{StringId, TypeId};
 use super::instructions::{Call, Match, Opcode, Return, Trampoline};
@@ -31,28 +32,69 @@ fn read_u32_le(bytes: &[u8], offset: usize) -> u32 {
     ])
 }
 
-/// Storage for bytecode bytes.
-#[derive(Debug)]
-pub struct ByteStorage(Vec<u8>);
+/// Storage for bytecode bytes with guaranteed 64-byte alignment.
+///
+/// All bytecode must be 64-byte aligned for DFA deserialization and cache
+/// efficiency. This enum ensures alignment through two paths:
+/// - `Static`: Pre-aligned via `include_query_aligned!` macro
+/// - `Aligned`: Allocated with 64-byte alignment via `AlignedVec`
+pub enum ByteStorage {
+    /// Static bytes from `include_query_aligned!` (zero-copy, pre-aligned).
+    Static(&'static [u8]),
+    /// Owned bytes with guaranteed 64-byte alignment.
+    Aligned(AlignedVec),
+}
 
 impl Deref for ByteStorage {
     type Target = [u8];
 
     fn deref(&self) -> &Self::Target {
-        &self.0
+        match self {
+            ByteStorage::Static(s) => s,
+            ByteStorage::Aligned(v) => v,
+        }
+    }
+}
+
+impl std::fmt::Debug for ByteStorage {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            ByteStorage::Static(s) => f.debug_tuple("Static").field(&s.len()).finish(),
+            ByteStorage::Aligned(v) => f.debug_tuple("Aligned").field(&v.len()).finish(),
+        }
     }
 }
 
 impl ByteStorage {
-    /// Create from owned bytes.
-    pub fn from_vec(bytes: Vec<u8>) -> Self {
-        Self(bytes)
+    /// Create from static bytes (zero-copy).
+    ///
+    /// The bytes must be 64-byte aligned. Use `include_query_aligned!` macro.
+    ///
+    /// # Panics
+    /// Panics if bytes are not 64-byte aligned.
+    pub fn from_static(bytes: &'static [u8]) -> Self {
+        assert!(
+            (bytes.as_ptr() as usize).is_multiple_of(64),
+            "static bytes must be 64-byte aligned; use include_query_aligned! macro"
+        );
+        Self::Static(bytes)
+    }
+
+    /// Create from an aligned vector (from compiler or file read).
+    pub fn from_aligned(vec: AlignedVec) -> Self {
+        Self::Aligned(vec)
+    }
+
+    /// Create by copying bytes into aligned storage.
+    ///
+    /// Use this when receiving bytes from unknown sources (e.g., network).
+    pub fn copy_from_slice(bytes: &[u8]) -> Self {
+        Self::Aligned(AlignedVec::copy_from_slice(bytes))
     }
 
-    /// Read a file into memory.
+    /// Read a file into aligned storage.
     pub fn from_file(path: impl AsRef<Path>) -> io::Result<Self> {
-        let bytes = std::fs::read(path)?;
-        Ok(Self(bytes))
+        Ok(Self::Aligned(AlignedVec::from_file(path)?))
     }
 }
 
@@ -118,15 +160,46 @@ pub struct Module {
 }
 
 impl Module {
-    /// Load a module from owned bytes.
-    pub fn from_bytes(bytes: Vec<u8>) -> Result<Self, ModuleError> {
-        Self::from_storage(ByteStorage::from_vec(bytes))
+    /// Load a module from an aligned vector (compiler output).
+    ///
+    /// This is the primary constructor for bytecode produced by the compiler.
+    pub fn from_aligned(vec: AlignedVec) -> Result<Self, ModuleError> {
+        Self::from_storage(ByteStorage::from_aligned(vec))
+    }
+
+    /// Load a module from static bytes (zero-copy).
+    ///
+    /// Use with `include_query_aligned!` to embed aligned bytecode:
+    /// ```ignore
+    /// use plotnik_lib::include_query_aligned;
+    ///
+    /// let module = Module::from_static(include_query_aligned!("query.ptk.bin"))?;
+    /// ```
+    ///
+    /// # Panics
+    /// Panics if bytes are not 64-byte aligned.
+    pub fn from_static(bytes: &'static [u8]) -> Result<Self, ModuleError> {
+        Self::from_storage(ByteStorage::from_static(bytes))
     }
 
     /// Load a module from a file path.
+    ///
+    /// Reads the file into 64-byte aligned storage.
     pub fn from_path(path: impl AsRef<Path>) -> Result<Self, ModuleError> {
-        let storage = ByteStorage::from_file(&path)?;
-        Self::from_storage(storage)
+        Self::from_storage(ByteStorage::from_file(&path)?)
+    }
+
+    /// Load a module from arbitrary bytes (copies into aligned storage).
+    ///
+    /// Use this for bytes from unknown sources (network, etc.). Always copies.
+    pub fn load(bytes: &[u8]) -> Result<Self, ModuleError> {
+        Self::from_storage(ByteStorage::copy_from_slice(bytes))
+    }
+
+    /// Load a module from owned bytes (copies into aligned storage).
+    #[deprecated(since = "0.1.0", note = "use `Module::from_aligned` for AlignedVec or `Module::load` for copying")]
+    pub fn from_bytes(bytes: Vec<u8>) -> Result<Self, ModuleError> {
+        Self::load(&bytes)
     }
 
     /// Load a module from storage.